IO-LVM

This project is related to the NeurIPS 2025 paper "Inverse Optimization Latent Variable Models for Learning Costs Applied to Route Problems".

Main files for training and saving models:

main_spp_synthetic.py

Related to the Waxman generated graphs and paths experiment. Here you can set the input "mult" to 1 or to 0 to choose between the multiple start/target nodes versus the single start/target nodes experiment.

main_spp_cabspot.py

Related to the taxi trajectories experiment. Please extract the data found in "cabspotting_preprocessing" folder.

main_spp_ship.py

Related to the ship trajectories experiment.

main_tsp_tsplib.py

Related to the tsp_lib experiment with Hamiltonian cycles generated data.

Main Arguments:

1. latent_dim: Number of latent dimensions, check the paper for a better grasp on values to choose.
2. method: Most important values are "IOLVM" and "VAE".
3. alpha_kl or beta: KL regularization. Depending on the experiment you might correct the value according to the batch size (e.g., whatever is given in the paper/BS).
4. n_epochs: Number of epochs.
5. eps: The perturbation for gradient estimation, 0.05 generally works fine.
6. lr: Learning Rate, defaults work fine.

Jupyter file to explore saved models (synthetic example)

compare_synthetic_eval.ipynb

Related to the Waxman generated graphs and paths experiment. I saved a IOLVM .pkl and a VAE .pkl for a comparison purpose, but feel free to train with different parameters and try it yourself.

---